低分辨率条件下基于TLD的鲁棒车辆跟踪算法

摘要

TLD (Tracking-Learning-Detection)is a widely adopted algorithm in the research topics of vehicle tracking.However,the tradi-tional TLD is easily to lose targets and locate the wrong targets in a low-resolution video,and that is why the traditional TLD algorithm should be improved.As the pyramid optical flow tracker in traditional TLD fails to work well in poor illumination scenes,a novel vehicle tracking algo-rithm by applying the CT(Compressive Tracking)into the traditional TLD algorithm is proposed to obtain the vehicle location fast when the targets are lost.Besides,with the multi-features fusion of the 2bitBP-HOG feature descriptor,which is composed of shape and texture informa-tion in tracking phase,the improved algorithm obtains higher-performance tracking results than those of traditional TLD in low-resolution con-dition,not only tracking precision but also time complexity.Experiment results validate the improvement of the proposed algorithm,both on ro-bustness and tracking rates.%TLD（Tracking-Learning-Detection）算法是一种广泛应用的车辆跟踪算法，但其在低分辨率视频下易出现跟踪漂移或者丢失等问题，为此，对传统的 TLD 算法进行改进。针对 TLD 算法中金字塔光流跟踪器易受光照影响，出现跟踪失败问题，采用具有较强跟踪性能的 CT 跟踪器，并研究跟踪失败自检测策略，以提高算法的跟踪性能。此外，通过对2bitBP-HOG 特征（形状和纹理特征）描述算子进行多特征融合，有效克服了低分辨率环境下纹理特征提取不准确造成检测器准确度低的问题。实验表明，改进算法在鲁棒性和跟踪速率方面都有所提高。